In recent years a great amount of research has focused on algorithms that learn features from unlabeled data. These approaches are known as feature learning or deep learning methods and have been successfully applied to classify scene images and recognize with high precision handwritten characters. In this thesis we show that a feature learning approach can be used to segment complex textures, a problem for a long time addressed proposing a large amount of handcrafted descriptors and local optimization strategies. We employ the SOM neural network for its ability to natively provide a set of topologically ordered features. These features allow us to obtain a highly accurate local description, even in areas characterized by a transition from one texture to another. We also show that a single feature learning unit can be combined with others in order to significantly improve the quality of the texture description and, consequently, reduce the segmentation errors. The results obtained prove that the proposed segmentation method is valid and provides a real alternative to other state-of-the-art methods. Since the proposed framework is simple, we easily combined it with a pyramidal histogram encoding and a SVM supervised network in order to classify scene images. We show that the important topological ordering property, inherited from the SOM network, allow us to resize the feature set, obtained during the initial unsupervised learning, avoiding an unpredictable performance loss. Moreover, the results on the standard Caltech-101 dataset prove a significant improvement on some state-of-the-art computer vision methods, designed specifically for image classification.
展开▼
